Fast-COCO-Eval

This package wraps a facebook C++ implementation of COCO-eval operations found in the pycocotools package. This implementation greatly speeds up the evaluation time for coco's AP metrics, especially when dealing with a high number of instances in an image.

Comparison

For our use case with a test dataset of 1500 images that contains up to 2000 instances per image we saw up to a 100x faster evaluation using fast-coco-eval (FCE) compared to the original pycocotools code.

Seg eval pycocotools 4 hours 
Seg eval FCE: 2.5 min

BBox eval pycocotools: 4 hours 
BBox eval FCE: 2 min

Getting started

Install

pip install fast-coco-eval

If you clone the repo and install it locally, the following command is recommended

pip install -e .

given that you are in the fast-coco-eval directory. There seem to be an issue with loading the C++ extensions when installing it from the root directory without the -e flag.

Usage

This package contains a faster implementation of the pycocotools COCOEval class. Due to torch being used to compile and access the C++ code, it needs to be imported before using the package. To import and use COCOeval_fast type:

import torch
from fast_coco_eval import COCOeval_fast

For usage, look at the original COCOEval class documentation.

Dependencies

• pytorch>=1.5
• pycocotools
• pybind11
• numpy

It would be nice to decouple it from the pytorch build tool for the c++ compilation.

TODOs

• Wrap c++ code
• Get it to compile
• Add COCOEval class wraper
• Remove detectron2 dependencies
• Check if it works on windows
• Remove torch dependencies

License

Distributed under the apache version 2.0 license, see license for more information. © 2021 Sartorius AG